Open yifanfeng97 opened 6 years ago
I was also wandering. In the paper, the author use VGG-M and VGG-VD. So why many people implement it using AlexNet?(the Caffe implementation uses AlexNet as well)
@youkaichao
My reason is I had difficulty finding usable pretrained weight for VGG-M (and can be converted for Tensorflow) when I began this project 😅.
(VGG-M was implemented in matconvnet)
@XDUfyf I haven't tried VGG19. Did you try reducing the batch size?
@WeiTang114 I have tried reducing the batch size, and I set batch size with 1. It did not work, too.
At the same time, I met a strange problem. when I tried train with single view not multi-view(based on VGG19, my multi-view didn't work. so I temporarily use single view as input ) in ModelNet40. If I set batch size as 64 ,training would get nan in step 404. however I set batch size as 16, training would get nan in step 1619( 64 = 164 and 1619 = 4044). No matter how many times I run. It would get same result. My loss curve is normal. environment: Tensorflow, Python3, TITAN X , Cuda 8.
The resnet 101 is definitely a good choice to try for better performance.
Hi, Can we train our model from scratch without the pretrained AlexNet model? Thanks!
I used Tensorflow, and i was besed on VGG19. But I just got the OOM error! I had reused the variable in the loop! I used TITAN X.